We propose a five-year effort to extend the limits of behavioral methodology for teaching individuals with mental retardation and other severe learning difficulties. These individuals typically have communication skill limitations that render them difficult or impossible to teach with traditional methods of instruction. Special education has the goal of preparing these individuals for placement and increased participation in the community, rather than in institutions or other special settings. However, skills acquired in special education settings are broadly useful only if they generalize or transfer to other environments. Regrettably, such skills are frequently under narrow stimulus control, that is, those skills are not displayed outside the situations in which they were taught. Our proposed project will study problems of generalization and transfer in the context of a broader program that seeks to apply computer technology to teach rudimentary reading skills to individuals with severe and moderate intellectual disabilities. In research already accomplished, we have developed an effective methodology for engendering complex, integrated networks of matching, naming, and writing skills. Our research program and its general methodology provide a natural framework within which to study problems of generalization and transfer. Clearly, rudimentary reading and related skills acquired via computer teaching will not be functional unless they transfer to other settings. We take a novel stimulus-class approach for the experimental analysis of the problems of generalization and transfer. A stimulus class is a set of stimuli that are functionally equivalent in their control of behavior. The functional equivalence may result because (a) the stimuli share physical characteristics, and hence are in the same feature stimulus class (e.g., printed words stop and stop being class-members)or (b) training establishes arbitrary stimulus relations (e.g., between spoken and printed words "STOP" and STOP which yield an arbitrary stimulus class. Resembling prior conceptual analyses, our approach suggests that the formation of feature classes constitutes an important basis for cross-situational transfer and generalization. Unique to our analysis, however, is an emphasis on relationships between feature classes and arbitrary classes which foster the transfer of stimulus control across situations. The studies proposed here begin to explore systematically implications and parameters of our stimulus class model.